Just another WordPress site - Ruhr-Universität Bochum

Department Atomistic Modelling and Simulation
Research Group

Data-Driven Methods for Atomistic Simulations

The research group develops and applies data-driven methods in materials science, with a principal emphasis on atomic-scale simulations using the Atomic Cluster Expansion (ACE) – a new type of machine learning interatomic potentials with a formally complete basis set.

RUB, Marquard

Dr. Yury Lysogorskiy

Research Group Leader

Room: 02-719
Tel.: +49 234 32 29300
E-Mail: yury.lysogorskiy@icams.rub.de

Research

Our research covers the full cycle of ACE model parameterization and validation. This comprises extensions to the formalism, implementation in high-performance simulation codes, such as LAMMPS, parameterization of ACE using non-linear optimization with TensorFlow, uncertainty indication and active learning for selecting representative data. It also includes deploying high-throughput calculations for computing reference DFT energies and forces and workflows for validating interatomic potentials for accuracy and transferability.

Block scheme of the main pacemaker workflow.

ICAMS, RUB

Competences

Atomic Cluster Expansion (ACE): method development, parameterization and validation
High-throughput calculations (DFT and molecular dynamics)
Data-driven methods in materials science: machine learning, generative models

Members

Recent Publications

Research Examples

Multilayer atomic cluster expansion for semi-local interactions

The multilayer atomic cluster expansion (ml-ACE) was presented, which includes collective, semi-local multiatom interactions naturally within its remit. It was demonstrated that ml-ACE significantly improves fit accuracy and efficiency compared to a local expansion on selected examples and provides physical intuition to understand this improvement.

Performant implementation of the atomic cluster expansion (PACE): application to copper and silicon

We implemented the ACE in the performant C++ code PACE that is suitable for use in large-scale atomistic simulations with LAMMPS. It was demonstrated that the atomic cluster expansion as implemented in PACE shifts a previously established Pareto front for machine learning interatomic potentials toward faster and more accurate calculations.